FairnessPrivacyBiascs.SI
Fairness Through Controlled (Un)Awareness in Node Embeddings
Dennis Vetter, Jasper Forth, Gemma Roig, Holger Dell
July 29, 2024
arXiv.org
Key Takeaways
-
Graph algorithms can be tuned to hide or reveal sensitive information: The paper demonstrates that by adjusting parameters in the CrossWalk algorithm, researchers can control whether sensitive attributes (like protected characteristics) remain detectable in graph embeddings, offering a practical lever for fairness control.
-
One-size-fits-all fairness doesn't work: The research shows that different fairness approaches may require different levels of attribute detectability, suggesting that organizations need flexibility to adapt graph algorithms to their specific fairness goals and regulatory contexts.
-
Fairness in AI requires intentional design choices: The findings indicate that fairness outcomes in graph-based machine learning systems are not automatic—they result from deliberate hyperparameter decisions, emphasizing that responsible AI requires active technical choices rather than passive hope.
-
Social networks and graph data pose unique fairness challenges: Because graph representation learning is "central" to applications like social networks that involve "sensitive personal data," this research addresses a high-stakes domain where fairness failures could have direct societal implications for real people.
Abstract
Graph representation learning is central for the application of machine learning (ML) models to complex graphs, such as social networks. Ensuring `fair' representations is essential, due to the societal implications and the use of sensitive personal data. In this paper, we demonstrate how the parametrization of the \emph{CrossWalk} algorithm influences the ability to infer a sensitive attributes from node embeddings. By fine-tuning hyperparameters, we show that it is possible to either significantly enhance or obscure the detectability of these attributes. This functionality offers a valuable tool for improving the fairness of ML systems utilizing graph embeddings, making them adaptable to different fairness paradigms.